Licence: the decoder is a Qwen2.5-3B-Instruct derivative (Qwen Research License, non-commercial) — details and asks in GitHub issue #2
Thanks for the model — we evaluated it this week for an Apple Silicon (MLX / Swift) port into a commercial dictation app and like the design a lot. One licence point, written up with permalinks in https://github.com/Edge0-AI/Audio8-ASR-Infinite/issues/2:
The card says Apache-2.0, but the README's architecture table and config.json (text_config._name_or_path: "Qwen/Qwen2.5-3B-Instruct", Qwen2.5-3B's exact geometry, Qwen2.5's chat template verbatim) say the decoder and LM head were initialised from Qwen/Qwen2.5-3B-Instruct — and unlike the 0.5B / 1.5B / 7B / 14B / 32B sizes, Qwen2.5-3B is under the Qwen RESEARCH LICENSE AGREEMENT: derivative works are licensed "FOR NON-COMMERCIAL PURPOSES ONLY" (§2.a), commercial use needs a licence from Alibaba (§2.b), and §3.d lets a licensee relicense only their own additions. As far as we can read it, that leaves the ~3.09 B decoder parameters non-commercial for downstream users, whatever the badge says. The §3.a / §3.c notices and the §4.b "Built with Qwen" line are also missing from this repo.
Asks (details in the issue): (1) a variant with the decoder retrained from an Apache-2.0 base — Qwen2.5-1.5B-Instruct, Qwen3-1.7B / 4B, or Ministral-3-3B (Voxtral Realtime's own decoder); a smaller one would also help on-device; (2) a licence note on this card naming the decoder's base and its licence, plus the notices; (3) if there is an arrangement with Alibaba that covers downstream commercial use, a line on the card saying so.
Happy to be corrected on the reading, and happy to test a permissive-base checkpoint on Apple Silicon when one exists.